Impact of Atmospheric Correction on Machine Learning-Based Land Cover Classification: Insights from the Sil Canyon (NW Spain)

• Sen2Cor, MAJA and iCOR corrections are used on Sentinel-2 images. • RF, SVM and CNN algorithms are compared. • Accuracies above 80 % are obtained for the CNN algorithm. • The results obtained show no significant differences in the final classifications. • The methodological processes used can be applied to other areas of study. • All results contribute to the management of protected areas. The popularisation of remote sensing systems (RS), satellite missions and open-access tools has facilitated the use of multiple platforms and algorithms to classify land cover. This fosters the development of cost-effective and sustainable mapping approaches. This study presents a unified methodological framework that systematically evaluates atmospheric correction (AC) strategies and classifier robustness in object-based (OBIA) land cover mapping. Three AC methods (Sen2Cor, MAJA and iCOR) were compared to assess their influence on land cover classification performance, a key factor for urban planning and natural resource management. The performance of these corrections was assessed in combination with three widely used classification algorithms (Random Forest, RF; Support Vector Machine, SVM; and Convolutional Neural Networks, CNN). The Sil Canyon in Galicia (NW Spain) served as the study area. Its steep topography and limited accessibility make extensive fieldwork unfeasible. Therefore, Sentinel-2 Level-1C multispectral spring imagery (2016–2023) was selected and atmospherically processed to surface reflectance products (Level-2A or equivalent). This selection aimed to enhance temporal consistency and minimize inter-annual variations related to vegetation phenology and illumination conditions. OBIA was used for classification. The results revealed a clear distinction between the best and the most robust performances. Statistical analysis (Kruskal-Wallis) confirmed that the choice of AC method had no significant effect on accuracy ( p = 0.6053), while no classifier was statistically superior in median performance ( p = 0.92). Rather than identifying a superior atmospheric correction approach, the results demonstrate that classifier robustness is the most influential factor for achieving reliable land cover classification, outweighing the choice of AC method. The CNN, for example, proved highly unstable, achieving both the highest overall accuracy (OA = 90.4 %) for the 2023 image and the lowest OA (50.4 %) for 2016. In contrast, Random Forest (RF) demonstrated the most robust performance, showing the least variability (OA Standard Deviation (SD) = 8.78 %). The open-source workflow developed in this study can support knowledge-based decision-making aligned with the Sustainable Development Goals.